Lumen-Rec
Controlled comparison of pointwise LLM relevance estimates and uncertainty-aware reranking.
Research question
LLM recommender comparisons often change the candidate set, model backbone, importer, and metric at the same time. Lumen-Rec asks what remains when those protocol variables are held fixed.
Approach
The repository fixes a shared 101-item candidate task, a Qwen3-8B backbone, a score schema, an evaluator, provenance records, and paired statistics. Within that protocol it compares a pointwise relevance posterior with official-code-level baselines and uncertainty-adjusted variants.
Evidence and boundaries
The repository reports that its posterior ranks first in six of eight evaluated domains, while retaining losses on Beauty and Movies; it also reports that the added uncertainty/risk decomposition does not improve this same-candidate ranking. These numbers are repository-reported evidence, not an independent replication by this site. The result is limited to same-candidate reranking: it is not a full-catalog recommendation result, a universal state-of-the-art claim, or a calibration guarantee.
Artifacts and provenance
- Source code and research record
- Heatmap figure source
- Observation figure source
- Project code and generated figures are distributed under the repository’s MIT license. Baseline and dataset provenance is retained in the repository rather than re-authored here.